Quantcast
View Post
drkohler said:

While that sounds like a good idea at first, it really isn't practical for something like a console which has to be cheap to manufacture.

The whole idea of the chiplet design is to reduce manufacturing costs.
Things like I/O, memory controllers and so on... Don't scale down in manufacturing nodes very well... So it makes sense to keep them on an older, more mature process that is cheaper.

drkohler said:

Chiplet designs require much more time to design, it requires interfaces between the chiplets (and this at very high speeds) more testing individual components, and some other, more obscure things.

AMD has a team dedicated to that very task.
But now that Infinity Fabric is a known quantity... It's very easy for AMD to build that out and take full advantage of it very quickly.

drkohler said:

 In the end, the price will be too high so I expect a monolithic chip, again. One chip still is the most cost effective solution for products that have to be cheap to manufacture.

I am not willing to make an assumption either way until I have more information.

Trumpstyle said:

There was recently an interview with a AMD rep and two question were interesting for us regarding next-gen console. It looks like ps5 and Next Xbox will have a chiplet design instead of a SOC/APU.

It's a "possibility". - Awhile ago there was a thread where people were making guesses at what the next gen consoles would have... And Ryzen+Infinity Fabric was mentioned as one approach that could be taken.

The big advantage is scalability... Microsoft and Sony could with relative ease build an entire lineup of consoles with differing GPU capabilities whilst keeping the CPU+Chipset+I/O+Ram identical.


Trumpstyle said:

IC: With chiplets connected via IF on Rome, if a customer wanted a semi-custom design with different IP, such as a GPU or an AI block or an FPGA, would that be possible? (Say for example, a console?)

MP: Our semi-custom group is wide open to talk to customers to brainstorm! What excites me about the chiplet approach is that I think it’s going to disrupt the industry. It’s going to change the way the industry dreams of different configurations. Some might be right, and I can guarantee that someone will conjure up ten other ones that we didn’t think of! Honestly I think it is a disruptive force that is just nascent, just starting right now.

This was always the case, Anandtech just got clarification.

Trumpstyle said:

IC: With IF on 7nm, it offers 100 GB/s GPU to GPU connectivity. One of your competitors has something similar which allows both GPU-GPU and CPU-GPU connectivity. Currently with Rome, PCIe 4.0 has been announced from CPU to GPU but not IF. What has AMD’s analysis been on that CPU to GPU link?

MP: We haven’t announced applying the IF between the CPU and GPU and while it is certainly feasible, it is likely just dependent when workloads could truly leverage that protocol being applied, when the full coherency is required across both CPU and GPU. It is certainly feasible, but we haven’t announced it at this time.

Obviously playing coy. I don't see no reason why that approach couldn't be taken.


Trumpstyle said:
Basically what this means is that Sony and Microsoft can pump in more Gpu cores into their next console. Something like a 88CU clocked at 1400mhz giving 15,7 Teraflops seems very feasible. Probably Vivster or Pemalite can give a better explanation what chiplet design is.

Because the individual chips are smaller, they can get better yields, which means lower costs on the latest manufacturing process.
So in theory it should mean more CU's.